Will AI agents replace SaaS tools? Partly, and not the part most people expect. Agents are already absorbing the thin glue layer of software, the point tools, single-purpose automations, and light CRUD apps that exist mainly to move a record from A to B. But systems of record and deep domain platforms are not going anywhere soon, and the honest near-term reality is that agents sit on top of your existing SaaS through APIs rather than ripping it out.

I run Gravity, an AI agent platform, so I have an obvious bias toward "agents win." I am going to argue against my own interest where the evidence demands it. The "agents eat SaaS" thesis is half right and gets oversold by the other half. This piece separates the durable claim from the hype, names what genuinely gets disrupted versus what survives, and gives you a timeline you can actually plan against.

Key takeaways

  • AI agents will absorb the thin, workflow-glue layer of SaaS: point tools, single-purpose automations, and light CRUD apps. That part of the market is genuinely at risk.
  • Systems of record and deep domain platforms endure. The hard part of enterprise software was never the UI; it was the data model, compliance, and integrations underneath.
  • The near-term reality is agents sitting on top of SaaS through APIs, not replacing it wholesale. Your CRM stays; the agent reads and writes to it for you.
  • The pricing model is what actually breaks first. Per-seat SaaS economics struggle when one agent does the work of ten seats; per-use pricing fits agents far better.
  • Honest timeline: glue-layer disruption is already happening in 2026, deeper platform displacement is a multi-year story, and most of it is hype until the agent reliably finishes the job.
The short answer
The short answer

The short answer

Software does not collapse into a single agent overnight. What happens is narrower and more useful to understand: the value that lived in workflow plumbing migrates to a reasoning layer, while the value that lives in data, compliance, and deep domain logic stays put.

Think of any SaaS product as two things stacked together. Underneath sits a system of record, a structured database with hard-won rules about how your customers, invoices, or inventory behave. On top sits an interface and a pile of workflow logic, the forms, buttons, and "if this then that" automations that let a human drive the database. Agents are very good at the top layer and weak at safely owning the bottom one. So the top layer is what gets eaten first.

If your product is mostly the top layer, a glorified form over a table, you should be nervous. If your product is the table itself, the canonical record other systems trust, you have more runway than the headlines suggest. The interesting fights happen in the middle, where a vendor owns both and an agent platform tries to slide between them.

What the discourse actually says

The loudest version of this thesis came from Satya Nadella on the BG2 podcast with Bill Gurley and Brad Gerstner in May 2025. His framing was blunt: "the notion that business applications exist, that's probably where they'll all collapse, right in the agent era, because... they are essentially CRUD databases with a bunch of business logic. The business logic is all going to these agents." You can read the full quote and context here. Microsoft's own Dynamics roadmap is moving in that direction, so it is not idle speculation.

It is worth reading that quote carefully, because it concedes the exact split I am drawing. Nadella says the business logic moves to agents. He does not say the CRUD databases vanish. The records still exist; something has to be the source of truth. What collapses is the bespoke logic and UI layer that vendors spent two decades building on top of those records.

On the investor side, a16z has been making a quieter and, to me, sharper version of the argument. Their December 2024 enterprise note put it plainly: "Per-seat is no longer the atomic unit of software." As they describe it, when AI handles the actual work, the value metric shifts from how many humans log in to how much output gets delivered. You can see the pricing argument in full here. That is the part of the thesis I think is most durable, and I will come back to it.

The counter-discourse matters too. Plenty of operators point out that ripping out a system of record is a multi-year, career-risking project, and that an agent that occasionally hallucinates a database write is not allowed anywhere near production finance data. Both camps are right about different layers. The mistake is treating "SaaS" as one monolithic thing that either dies or doesn't.

What gets disrupted vs what survives

Here is my actual scorecard, drawn from building agents that have to interoperate with real customer stacks. It is not a clean line, but the pattern is consistent.

What gets disrupted, and fast:

What survives, and arguably gets stronger:

The honest uncomfortable middle is the mid-market SaaS suite that is mostly workflow with a shallow data model. Those vendors face a real choice: become the system of record their customers trust, or become a feature an agent calls. Many will not make that jump.

Agents sit on top of SaaS, for now

Strip away the rhetoric and look at how agents actually ship in 2026. They do not replace Salesforce. They log in to Salesforce, read the pipeline, draft the follow-ups, update the records, and hand a summary back. The SaaS tool is still there. The agent is a new layer of intent sitting above it, talking to it through APIs and increasingly through shared protocols.

This is not a temporary embarrassment for the thesis; it is the rational path. APIs are how agents get reliable, permissioned access to systems of record without reimplementing twenty years of domain logic. The emergence of interoperability standards, which I track in our coverage of AI agent future trends for 2026, makes the on-top pattern stronger, not weaker, because it gives agents a sanctioned, auditable way in.

It also explains why the most successful early agents look like better front doors to existing tools rather than replacements. They collapse the human workflow without collapsing the underlying system. That is the same reason teams reach for an agent platform instead of stitching together brittle automations: you can see the difference in practice in the best Zapier alternatives using AI and in our head-to-head on Gravity versus Zapier. The agent absorbs the glue; the apps it touches keep doing what they do well.

Over a longer horizon some of those underlying tools do get hollowed out, especially the thin ones. But "agent calls your stack today, agent slowly absorbs the weakest parts of your stack over years" is a very different prediction from "SaaS is dead." Only one of those is happening right now.

Describe the outcome, not the workflow

The deepest reason agents eat the glue layer is a change in what users have to specify. Traditional SaaS forces you to translate your goal into the software's vocabulary: pick this menu, configure that rule, map these fields, build the seven-step automation. You do the translation work. The tool just executes the steps you already designed.

Agents flip that. You state the outcome, "reconcile last month's invoices against the bank feed and flag anything over a thousand dollars that doesn't match," and the agent figures out the steps. This is the core of a thesis I have written about at length in describe the outcome, not the workflow, and it is why the workflow layer is the first to go. Once a machine can plan the steps, the value of a product whose entire job was to let humans hand-build those steps drops sharply.

This is also where the "agent" distinction actually matters and is not just branding. A workflow tool executes a fixed graph you authored. An agent reasons about a goal, chooses tools, handles the branches you did not anticipate, and recovers from failures. That capability gap is the whole ballgame, and it is why I am bullish on the category even while I am skeptical of the "SaaS is dead" headline.

It does not eliminate the systems underneath. Reconciling invoices still requires a ledger and a bank feed that are correct and trusted. The agent describes-the-outcome on top; the systems of record keep being right underneath. That division of labor is the realistic shape of the next few years, and it is exactly what use cases like the ones in the top AI agent use cases for H1 2026 demonstrate.

The pricing shift: per-seat to per-use

If you want to know which part of SaaS breaks first, follow the money, not the technology. The pricing model cracks before the product does.

Per-seat pricing assumes a human sits in a chair and uses the software. That assumption is the entire business model for most SaaS companies. Now run the agent scenario: one agent does the work of several seats, around the clock, without logging in as ten different users. The seat-based meter no longer maps to the value delivered, and as a16z argued, "per-seat is no longer the atomic unit of software." Their canonical example is customer support, where buyers historically paid roughly $115 per agent seat per month but the natural unit becomes a resolved ticket once AI handles resolution.

This is why I think the pricing implication is the most concrete, near-term consequence of the whole debate. Agents push pricing toward usage and outcomes: pay for the work done, not the chairs filled. At Gravity we went all-in on this early, billing per use rather than per seat, where a dollar buys a thousand credits and you pay for what an agent actually runs. The reasoning is laid out in our 2026 trends piece, and it is not altruism; it is just the model that fits how agents create value.

The squeeze for incumbents is real. A SaaS vendor charging per seat watches an agent shrink the number of humans who need a seat, which shrinks revenue even if the underlying product is still used. Some respond by repricing to usage; some bolt an agent onto their own suite and reprice around outcomes; some get disintermediated by an agent platform that sits above them. The vendors most exposed are precisely the thin, workflow-heavy ones from the scorecard above.

A realistic timeline

Predictions are cheap, so let me anchor mine to data rather than vibes. Adoption is real and broad: McKinsey's late-2025 State of AI survey found 88 percent of organizations report regularly using AI in at least one business function, up from 78 percent a year earlier. But the same survey is a cold shower on the pace, with only about one-third of organizations scaling AI across the enterprise. Adoption is wide and shallow, which is exactly what you would expect during a glue-layer transition, not a wholesale replacement.

The agent-specific reality is even more sobering. Gartner forecasts that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. That is not a reason to dismiss the category; it is a reason to be precise about timing. The same firm projects that by 2028, around 15 percent of day-to-day work decisions will be made autonomously by agents, up from essentially zero in 2024. Meaningful, but a long way from "SaaS is gone."

So here is my honest timeline, hedged on purpose:

The single biggest variable is reliability. Until an agent finishes the job correctly the overwhelming majority of the time, with audit trails a compliance team accepts, the replacement story stays capped at low-stakes glue work. That is the gating factor, and it is why I keep telling people the near-term winner is "agents plus your existing SaaS," not "agents instead of it."

Frequently Asked Questions

Will AI agents replace SaaS tools completely?

No, not completely or soon. Agents are absorbing the workflow-glue layer: point tools, single-purpose automations, and light CRUD apps. Systems of record like CRMs, ERPs, and ledgers endure because they hold canonical data and deep domain logic that agents rely on rather than rebuild.

Are AI agents replacing SaaS tools right now in 2026?

Partly. In 2026 most agents run on top of existing SaaS through APIs, reading and writing records on your behalf rather than removing the underlying software. Thin automations and basic internal tools are being displaced, but core platforms remain in place and are often used more heavily.

Why do people say SaaS will collapse in the agent era?

The phrase traces to Satya Nadella, who argued on the BG2 podcast that business applications are essentially CRUD databases plus business logic, and that the logic moves to agents. Note the nuance: the logic and interface collapse, while the underlying databases and records persist as sources of truth.

What kind of SaaS is most at risk from AI agents?

Thin, workflow-heavy products with shallow data models are most exposed: point tools that copy fields between apps, single-purpose automations, light CRUD apps, and basic dashboards. If a product's entire value is letting humans hand-build steps a machine can now plan, it is in the blast radius.

How does agent pricing differ from per-seat SaaS pricing?

Per-seat pricing charges per human user, which breaks when one agent does the work of many seats. Agent platforms move toward usage or outcome-based pricing: you pay for work performed, not chairs filled. Gravity bills per use, where one dollar buys a thousand credits, matching cost to actual agent runs.

What is the realistic timeline for agents disrupting SaaS?

Glue-layer disruption is already underway through 2027, though Gartner expects over 40 percent of agent projects to be canceled in that window due to reliability and cost. Mid-market workflow suites face pressure toward 2030, while deep platforms and systems of record persist as services agents call.

The bottom line

"Will AI agents replace SaaS tools?" is the wrong question phrased as a yes or no. The accurate answer is that agents replace the workflow-and-interface layer of software while depending more heavily on the data-and-domain layer underneath. The glue gets eaten; the systems of record get stronger; and for the next several years the dominant pattern is agents sitting on top of your existing stack, not in place of it.

My advice if you are a buyer: do not rip out your CRM to chase a headline, but do question every thin tool whose only job is moving data around, and watch your per-seat contracts as agents change the math. If you are a SaaS builder, the survival move is to be the trusted record, not the convenient form on top of it. And if you want to feel the "describe the outcome, not the workflow" shift firsthand, that is exactly what Gravity is built to do: state the result you want and let an expert-built agent run it on top of the tools you already use, in about 60 seconds.

Sources